Imitation Accelerated Q-learning on a Simulated Formula Student Driverless Racecar

Abstract

In the international Formula Student competition, only a handful compete in the driverless category. Most of them using expensive hardware such as LIDAR’s. By leveraging reinforcement learning, a cheaper camera based system can be created .In order to train this system a simulator based on a fork of Microsoft’s AirSim by Formula Technion was used. A virtual replica of a Formula Student car designed for 2020 by Align Racing UiA, functioned as the test vehicle. In order to decrease the required training time, a pre-trained imitation learning network was used. This was implemented into a Deep Q-Learning network in four different methods. The most successful method was able to accelerate the learning process by 36%.